Banking Intent Classifier (RoBERTa + LoRA)

roberta-base fine-tuned with LoRA on Banking77 (77 customer-support intents). LoRA adapters have been merged into the base weights, so this loads like any standard AutoModelForSequenceClassification checkpoint.

• Accuracy: 92.05% | Macro F1: 92.05% | Weighted F1: 92.05%

Code, training notebook, and the PII-redaction layer used in front of this model: https://github.com/mkianih/ai-banking-intent-classifier

label_encoder.joblib (scikit-learn LabelEncoder) is included alongside the model weights to map class indices back to intent names.